What's for Dinner/Aaj Kya Banega? A 20B model trained to solve my Mom's kitchen crisis without the chatty fluff

A developer built "What's for Dinner?", an open-source meal-planning assistant that suggests up to three Indian home-cooked dishes from ingredients on hand, by fine-tuning the 20B open-weight gpt-oss-20b model with the Tinker Cookbook into an adapter called mom-chef-v1. The adapter, hosted on Tinker's cloud GPUs, was trained on a synthetic dataset to give atomic, decisive responses rather than chatty validation, and the whole fine-tuning run cost $0.16. The React and FastAPI app uses the Backboard API to remember recently eaten meals and is deployed as a single web service on Render.

This is a submission for the Hacktoberfest Weekend Challenge: Build for a Friend https://dev.to/challenges/hacktoberfest-weekend-2026-10-01 The most dangerous question in any Indian household: "Aaj khane mein kya banau?" What's for dinner? Obviously, most days I'll say, "Make whatever you want." And there's a good chance either Mom gets frustrated with that answer, or we're having the same dish for the fourth time that week. So I built What's for Dinner? for my Mom — and honestly, for the survival of the whole family especially me . It's an AI-powered meal-planning assistant that takes what's actually sitting in the fridge , using the names we naturally use at home — palak , paneer , bache hue chawal , etc. — and suggests at most 3 Indian home-cooked dishes that can actually be made. It also remembers what the family has eaten recently, so nobody has to suffer through Aloo Gobi three days in a row. But the most important feature isn't the recipes. It's knowing when to stop talking. Stop asking. Start cooking. The demo is running on Render's free tier, so the first request may take a little longer while the service wakes up. The entire project is open source: A full-stack application that suggests meals based on your available pantry items. It leverages the Tinker API to sample LLM-generated suggestions and uses the Backboard API to maintain a history of your recently selected meals, ensuring you don't get the same suggestion twice main.py /api/suggest for meal suggestions and /api/select to save your choice. frontend/ train/ train.jsonl . render.yaml & build.sh The React frontend and FastAPI backend are deployed together as a single Web Service on Render, making the deployment simple and helping me make the most of the Hacktoberfest credits. I wanted to build something fast, lightweight, culturally aware, and actually useful in a real Indian kitchen. I fine-tuned the 20B open-weight gpt-oss-20b model using the Tinker Cookbook. I trained an adapter called mom-chef-v1 on a synthetic dataset containing: The goal wasn't just to teach the model about Indian food. I wanted to teach it how to respond . The adapter is hosted on Tinker's cloud GPUs, so my application doesn't need to carry around ~43 GB of model weights. And the best part: The entire fine-tuning run cost me just $0.16. I integrated Backboard as the family's memory layer. It keeps track of recent meals and provides that context to the model before it makes a suggestion. This means the AI isn't only looking at what's in the fridge. It also knows what we've already eaten. So if someone says: "Aloo ki sabzi toh kal hi bani thi." The system can treat that as a real constraint instead of suggesting the same thing again. The React application is compiled into static assets and served directly through the FastAPI backend. Everything is deployed as a single Web Service on Render . The deployment configuration is included in the repository in render.yaml . This project started because I tried using generic AI models for my Mom. They understood Hinglish and local ingredients, but they completely failed to understand how she wanted the answer. A typical interaction looked like this: Mom: "Aaj kya banau? Aloo hai, paneer hai aur palak bhi hai." AI: "Aloo ki sabzi sounds wonderful " Mom: "Aloo ki sabzi toh kal hi bani thi, koi nahi khayega." AI: "Yes, you're absolutely right Since you've already had aloo yesterday, let's explore some other delicious alternatives..." Mom doesn't want a conversation. She doesn't want validation. She wants the AI to eliminate the bad option and give her a practical answer immediately. So, I designed What's for Dinner? around atomic, decisive responses . Instead of a chatty assistant, it just gives the options: Palak Paneer Paneer Paratha Kadhai Paneer No explanations. Just the answer. This is where open innovation changed the game. Instead of fighting a closed model's inherently chatty nature with massive, brittle system prompts, I fine-tuned an open-weight model on household constraints and the exact short, decisive answers my Mom expects. The goal wasn't just to build an AI that knows Indian food. The goal was to build an AI that knows when to stop talking. And thanks to open weights, I was able to build it for exactly $0.16 . I'm submitting What's for Dinner? for: Best Use of Tinker — Fine-tuned gpt-oss-20b into mom-chef-v1 for concise, culturally-aware Indian meal suggestions. Fine-tuning cost: $0.16. Best Use of Render — Deployed the complete React + FastAPI application as a single Web Service on Render. Backboard — Used Backboard as the family's memory layer to track recent meals and preferences.